This paper explores data-driven methods for quantifying and incorporating spatio-temporal contextual information in the
mapping of land cover change. In remote sensing, area classes of land cover are typically mapped via statistical
manipulation of feature-space measurement, e.g., reflectance data, and other ancillary data. Contextual information has
been known to have the potential of increasing the accuracy of land cover classification and change detection, on the
ground that land cover often exhibits spatial and temporal correlations and, as such, should be properly accommodated.
In Bayesian methods, a priori probabilities of class occurrences can be considered as contextual information, which are
combined with class-conditional probability densities to arrive at discriminant decisions with minimized
misclassification. These prior probabilities may be made to vary locally to honor variability in the strengths of spatial
dependence in class occurrences. For deriving local prior joint probabilities in land cover co-occurrences over time, a
modified Expectation and Maximization (EM) algorithm was developed, in which a local window size can be adjusted in
the light of spatial dependences inferred from class probability densities computed from spectral data. Empirical studies
were performed using bi-temporal Landsat TM image subsets in Wuhan, which confirmed the comparative benefits of
incorporating localized prior probabilities in land cover change detection.
|